CryoLithe: Rapid Cryo-ET Reconstruction via Transform-Localized Deep Learning

TL;DR

CryoLithe achieves denoising and missing wedge correction comparable or better than state-of-the-art self-supervised deep learning approaches such as Icecream, Cryo-CARE, IsoNet or DeepDeWedge, while being two orders of magnitude faster, to implement a local, memory-efficient reconstruction network.

Abstract

Cryo-electron tomography (cryo-ET) enables 3D visualization of cellular structures. Accurate reconstruction of high-resolution volumes is complicated by the very low signal-to-noise ratio and a restricted range of sample tilts. Recent self-supervised deep learning approaches, which post-process initial reconstructions by filtered backprojection (FBP), have significantly improved reconstruction quality with respect to signal processing iterative algorithms, but they are slow, taking dozens of hours for an expert to reconstruct a tomogram and demand large memory. We present CryoLithe, an end-to-end network that directly estimates the volume from an aligned tilt series. CryoLithe achieves denoising and missing wedge correction comparable or better than state-of-the-art self-supervised deep learning approaches such as Icecream, Cryo-CARE, IsoNet or DeepDeWedge, while being two orders of magnitude faster. To achieve this, we implement a local, memory-efficient reconstruction network. We demonstrate that leveraging transform-domain locality makes our network robust to distribution shifts, enabling effective supervised training and giving excellent results on real data$\unicode{x2013}$without retraining or fine-tuning. CryoLithe reconstructions facilitate downstream cryo-ET analysis, including segmentation and subtomogram averaging and is openly available: https://github.com/swing-research/CryoLithe.

Figures (16)

• Figure 1: CryoLithe is a supervised deep learning method that is able to reconstruct tomograms in a couple of minutes without any parameter tuning. Existing self-supervised deep learning methods may require from hours to day and the expertise for training one or several neural networks. Our proposed localized architecture, inspired by the tomography imaging operator, scales favorably with any GPU whatever its memory capacity.
• Figure 2: Orthogonal slices though a variety of tomograms reconstructed with CryoLithe, representing diverse biological contexts and imaging conditions. CryoLithe can process arbitrary aligned tilt series and produce denoised and missing wedge corrected tomograms in 4 to 8 minutes on a GeForce RTX 4090 GPU with 24 GB memory. The reconstructed tomograms are cropped to form a square image on the $x-y$ axis. Table \ref{['table:mosaic-2']} provides information about the original data and their processing. The corresponding FBP reconstructions are displayed in Appendix, in Fig. \ref{['fig:mosaic-2-fbp']}.
• Figure 3: CryoLithe reaches similar FSC performance (above the 0.5-threshold) than state-of-the-art reconstruction algorithms on test data, but significantly faster (up to 75x) and without any parameter tuning. This makes CryoLithe an easy to use technique for high quality 3D reconstruction. FBP+ Icecream volume is used as reference to compute the FSC curves.
• Figure 4: Tomogram reconstruction from EMPIAR-12262 ishemgulova2024endosome tilt series containing non-infected cos-7 cells sampled at 5.525 Å/pixel. Projections are downsampled by a factor of 3. The tilt series is acquired between -51 and 51 degrees with a tilt increment of 3 degrees and with projections of size $1279 \times 1236$. We use AreTomo2 to align the projections before reconstruction. CryoLithe is able to process unseen dense biological structures on challenging datasets. Notice the processing time (up to $75$x faster) and the absence of expert intervention (no parameter to tune) for CryoLithe.
• Figure 5: Cross-correlation between the template of human 80S ribosome (EMD-2938 khatter2015structure) and the tomograms provided in the template matching tutorial dataset TLGJCM_2023. The tilt series are first pre-processed using Aretomo3 for alignment and CTF correction. The aligned tilt series serve as input to our method, while the corresponding FBP reconstruction from IMOD is used as input for FBP+Cryo-CARE+IsoNet and FBP+ Icecream. All volumes are reconstructed at Bin 4. We use pytom-match-pick chaillet2025pytom to perform template matching and obtain the cross-correlation scores. We then select the 200 particles with the highest scores and plot the values in descending order.